Closed TeaPoly closed 4 months ago
If the printed warnings are not a lot, you can just ignore it. The warnings just tell you that the prune_range
is too small to cover all the labels (transcript tokens) within t
steps, it will automaticaly increase the prune_range
. If the logs are a lot you can try using a larger prune_range
.
If the printed warnings are not a lot, you can just ignore it. The warnings just tell you that the
prune_range
is too small to cover all the labels (transcript tokens) withint
steps, it will automaticaly increase theprune_range
. If the logs are a lot you can try using a largerprune_range
.
Thanks for your reply.
The reason for I using extreme encoder output frame rate due to the paper Extreme Encoder Output Frame Rate Reduction: Improving Computational Latencies of Large End-to-End Models, which show that 320~5120ms output frame rate with HAT loss has comparable result by Alignment-Length Synchronous Decoding.
When I use encoder output frame rate with total 2560ms and
prune_range
is 5, there are some warning like this:When I setting
prune_range
to 20, there are other warning like this:What should I do to handle this problem?
This is how I use pruned rnn-t loss.